Sentence Analysis and Collocation Identification
Eric Wehrli, Violeta Seretan, Luka Nerima
Language Technology Laboratory
University of Geneva
{Eric.Wehrli, Violeta.Seretan, Luka.Nerima}@unige.ch
Abstract
the actual use made of these resources in other
NLP applications.
In this paper, we consider the particular application of syntactic parsing. Just as other types of
multi-word expressions (henceforth, MWEs), collocations are problematic for parsing because they
have to be recognised and treated as a whole, rather than compositionally, i.e., in a word by word
fashion (Sag et al., 2002). The standard approach
in dealing with MWEs in parsing is to apply
a “words-with-spaces” preprocessing step, which
marks the MWEs in the input sentence as units
which will later be integrated as single blocks in
the parse tree built during analysis.
We argue that such an approach, albeit sufficiently appropriate for some subtypes of MWEs2 ,
is not really adequate for processing collocations. Unlike other expressions that are fixed or
semi-fixed3 , collocations do not allow a “wordswith-spaces” treatment because they have a high
morpho-syntactic flexibility.
There is no systematic restriction, for instance,
on the number of forms a lexical item (such as a
verb) may have in a collocation, on the order of
items in a collocation, or on the number of words
that may intervene between these items. Collocations are situated at the intersection of lexicon
and grammar; therefore, they cannot be accounted
for merely by the lexical component of a parsing
system, but have to be integrated to the grammatical component as well, as the parser has to consi-
Identifying collocations in a sentence, in
order to ensure their proper processing in
subsequent applications, and performing
the syntactic analysis of the sentence are
interrelated processes. Syntactic information is crucial for detecting collocations,
and vice versa, collocational information
is useful for parsing. This article describes
an original approach in which collocations
are identified in a sentence as soon as possible during the analysis of that sentence,
rather than at the end of the analysis, as in
our previous work. In this way, priority is
given to parsing alternatives involving collocations, and collocational information
guide the parser through the maze of alternatives. This solution was shown to lead
to substantial improvements in the performance of both tasks (collocation identification and parsing), and in that of a subsequent task (machine translation).
1
Introduction
Collocations1 constitute a central language phenomenon and an impressive amount of work has
been devoted over the past decades to the automatic acquisition of collocational resources – as attested, among others, by initiatives like the MWE
2008 shared task aimed at creating a repository of
reference data (Grégoire et al., 2008). However,
little or no reference exist in the literature about
2
Sag et al. (2002) thoroughly discusses the extend to
which a “words-with-spaces” approach is appropriate for different kinds of MWEs.
3
For instance, compound words: by and large, ad hoc;
named entities: New York City; and non-decomposable
idioms: shoot the breeze.
1
We adopt the lexicographic understanding for the term
collocation (Benson et al., 1986), as opposed to the British
contextualist tradition focused on statistical co-occurrence
(Firth, 1957; Sinclair, 1991).
28
Proceedings of the Multiword Expressions: From Theory to Applications (MWE 2010), pages 28–36,
Beijing, August 2010
sing results. For instance, Brun (1998) compared
the coverage of a French parser with and without terminology recognition in the preprocessing stage. She found that the integration of 210
nominal terms in the preprocessing components of
the parser resulted in a significant reduction of the
number of alternative parses (from an average of
4.21 to 2.79). The eliminated parses were found
to be semantically undesirable. No valid analysis were ruled out. Similarly, Zhang and Kordoni (2006) extended a lexicon with 373 additional MWE lexical entries and obtained a significant
increase in the coverage of an English grammar
(14.4%, from 4.3% to 18.7%).
In the cases mentioned above, a “words-withspaces” approach was used. In contrast, Alegria et al. (2004) and Villavicencio et al. (2007)
adopted a compositional approach to the encoding of MWEs, able to capture more morphosyntactically flexible MWEs. Alegria et al. (2004)
showed that by using a MWE processor in the preprocessing stage of their parser (in development)
for Basque, a significant improvement in the POStagging precision is obtained. Villavicencio et al.
(2007) found that the addition of 21 new MWEs
to the lexicon led to a significant increase in the
grammar coverage (from 7.1% to 22.7%), without
altering the grammar accuracy.
An area of intensive research in parsing is
concerned with the use of lexical preferences, cooccurrence frequencies, collocations, and contextually similar words for PP attachment disambiguation. Thus, an important number of unsupervised (Hindle and Rooth, 1993; Ratnaparkhi, 1998;
Pantel and Lin, 2000), supervised (Alshawi and
Carter, 1994; Berthouzoz and Merlo, 1997), and
combined (Volk, 2002) methods have been developed to this end.
However, as Hindle and Rooth (1993) pointed
out, the parsers used by such methods lack precisely the kind of corpus-based information that
is required to resolve ambiguity, because many
of the existing attachments may be missing or
wrong. The current literature provides no indication about the manner in which this circular
problem can be circumvented, and on whether
flexible MWEs should be processed before, during or after the sentence analysis takes place.
der all the possible syntactic realisations of collocations.
Alternatively, a post-processing approach (such
as the one we pursued previously in Wehrli et
al. (2009b)) would identify collocations after the
syntactic analysis has been performed, and output a parse tree in which collocational relations
are highlighted between the composing items, in
order to inform the subsequent processing applications (e.g., a machine translation application).
Again, this solution is not fully appropriate, and
the reason lies with the important observation that
prior collocational knowledge is highly relevant
for parsing. Collocational restrictions are, along
with other types of information like selectional
preferences and subcategorization frames, a major
means of structural disambiguation. Collocational
relations between the words in a sentence proved
very helpful in selecting the most plausible among
all the possible parse trees for a sentence (Hindle
and Rooth, 1993; Alshawi and Carter, 1994; Berthouzoz and Merlo, 1997; Wehrli, 2000). Hence,
the question whether collocations should be identified in a sentence before or after parsing is not an
easy one. The previous literature on parsing and
collocations fails to provide insightful details on
how this circular issue is (or can be) solved.
In this paper, we argue that the identification of
collocations and the construction of a parse tree
are interrelated processes, that must be accounted
for simultaneously. We present a processing model in which collocations, if present in a lexicon,
are identified in the input sentence during the analysis of that sentence. At the same time, they are
used to rank competing parsing hypotheses.
The paper is organised as follows. Section 2
reviews the previous work on the interrelation
between parsing and processing of collocations
(or, more generally, MWEs). Section 3 introduces
our approach, and section 4 evaluates it by comparing it against the standard non-simultaneous approach. Section 5 provides concluding remarks
and presents directions for future work.
2
Related Work
Extending the lexical component of a parser with
MWEs was proved to contribute to a significant
improvement of the coverage and accuracy of par-
29
3
Parsing and Collocations
contribute to the disambiguation process so crucial for parsing. To put it differently, identifying
collocations should not be seen as a burden, as an
additional task the parser should perform, but on
the contrary as a process which may help the parser through the maze of alternatives. Collocations,
in their vast majority, are made of frequently used
terms, often highly ambiguous (e.g., break record,
loose change). Identifying them and giving them
high priority over alternatives is an efficient way
to reduce the ambiguity level. Ambiguity reduction through the identification of collocations is
not limited to lexical ambiguities, but also applies
to attachment ambiguities, and in particular to the
well-known problem of PP attachment. Consider
the following French examples in which the prepositions are highlighted:
As argued by many researchers – e.g., Heid (1994)
– collocation identification is best performed on
the basis of parsed material. This is due to the
fact that collocations are co-occurrences of lexical items in a specific syntactic configuration. The
collocation break record, for instance, is obtained
only in the configurations where break is a verb
whose direct object is (semantically) headed by
the lexical item record. In other words, the collocation is not defined in terms of linear proximity,
but in terms of a specific grammatical relation.
As the examples in this section show, the relative order of the two items is not relevant, nor is
the distance between the two terms, which is unlimited as long as the grammatical relation holds4 .
In our system, the grammatical relations are computed by a syntactic parser, namely, Fips (Wehrli, (1)a. ligne de partage des eaux (“watershed”)
2007; Wehrli and Nerima, 2009). Until now, the
b. système de gestion de base de données (“dacollocation identification process took place at the
tabase management system”)
end of the parse in a so-called “interpretation”
c. force de maintien de la paix (“peacekeeping
procedure applied to the complete parse trees. Alforce”)
though quite successful, this way of doing pred. organisation
de
protection
de
sents a major drawback: it happens too late to
l’environnement (“environmental protection
help the parser. This section discusses this point
agency”)
and describes the alternative that we are currently
developing, which consists in identifying colloca- In such cases, the identification of a nounpreposition-noun collocation will prevent or distions as soon as possible during the parse.
One of the major hurdles for non-deterministic courage any other type of prepositional attachparsers is the huge number of alternatives that ment that the parser would otherwise consider.
must be considered.
Given the high fre3.1 The Method
quency of lexical ambiguities, the high level of
non-determinism of natural language grammars, To fulfill the goal of interconnecting the parsing
grammar-based parsers are faced with a number procedure and the identification of collocations,
of alternatives which grows exponentially with the we have incorporated the collocation identificalength of the input sentence. Various methods tion mechanism within the constituent attachment
have been proposed to reduce that number, and procedure of our parser Fips (Wehrli, 2007). This
in most cases heuristics are added to the parsing parser, like many grammar-based parsers, uses
algorithm to limit the number of alternatives. Wi- left attachment and right attachment rules to build
thout such heuristics, the performance of a parser respectively left subconstituents and right submight not be satisfactory enough for large scale constituents. Given the fact that Fips’ rules always
applications such as machine translation or other involve exactly two constituents – see Wehrli
(2007) for details – it is easy to add to the attachtasks involving large corpora.
We would like to argue, along the lines of ment mechanism the task of collocation identificaprevious work (section 2), that collocations can tion. To take a very simple example, when the rule
attaching a prenominal adjective to a noun applies,
4
Goldman et al. (2001) report examples in which the disthe collocation identification procedure is invotance between the two terms of a collocation can exceed 30
ked. It first verifies that both terms bear the lexical
words.
30
feature [+partOfCollocation], which signals that a
given word is associated in our lexical database to
one or several collocations, and then searches the
collocation database for an adjective-noun collocation with those two terms. If successful, the corresponding parse tree will be given a high priority.
With examples such as loose change, the identification of the collocation will immediately relegate any (partial) analysis based on the verbal
reading of either terms.
To take a somewhat more complex example,
consider a verb-object collocation such as break
record, as in example (2)5 .
(2)a. John broke a record.
b. [
[ John ] broke [
TP
DP
DP
a [
NP
record ] ] ]
b. Jean a battu le détenteur du record.
“Jean has beaten the holder of the record”
As in the other examples, an analysis in which
a collocation has been found is given high priority over alternatives. In the case of (2), this will
relegate potential analyses based on the adjectival
reading of broke or the verbal reading of record.
Notice that exactly the same procedure applies
when the trace of an extraposed element is (right)
inserted, as in the examples (4), which illustrate
the case of wh-interrogative (a), relative clause
(b), tough-movement (c).
(4)a. Which record will Paul try to break ?
b. The record Paul broke was very old.
c. This record is difficult to break.
In all such cases, that is, when the right inserted
element is a trace, the identification procedure
will consider its antecedent, or to be more precise,
the semantic head of its antecedent. Finally, the
grammatical processes involved in example (4a,c)
can combine as in the more complex example (5),
for which we give the slightly simplified structure
with the chain of elements with index i extending
from the fronted wh-phrase which record to the
direct object position of the verb break, via the
direct object position of the verb consider and the
subject position of the secondary predicate (FP)
headed by the [+tough] adjective difficult.
Here, it is a right attachment rule which will
trigger the identification procedure. To be precise,
the right attachment rule in this case concerns the
attachment of the noun record as complement
of the indefinite determiner (head of the DP
direct object of the verb). The identification
procedure considers, in turn, all the governing
nodes dominating the noun record, halting at the
first node of category Noun6 , Verb or Adjective.
In our example, the determiner node and then
the verb node will be considered. Notice that the
procedure will, quite correctly, identify a collocation in the French example (3a), but not in (3b),
although both structures are identical. The reason
has to do with the fact that the noun governing (5)a. Which record did Paul consider difficult to
record in the first example is a [+number] noun,
break ?
that is a classifier noun which is transparent for
b. [ [ which record]i [ did [ Paul
the identification procedure7 .
CP DP
TP
DP
][
consider ][
e]i [
[
e]i [
VP
DP
FP DP
AP
(3)a. Jean a battu un grand nombre de records.
difficult [ to [ break [ e]i ] ] ] ] ] ]
TP
VP
DP
“Jean broke a large number of records”
5
We use the following labels in our phrase-structure representations: TP-Tense phrase, for simple sentence (the S of
standard CFG), CP-Complementizer phrase, for a sentence
with a conjunction or a complementizer, DP-Determiner
phrase for standard noun phrases (we assume the DP hypothesis, whereby the determiner constitutes the syntactic head of a noun phrase), NP-Noun phrase for nominal
projections (nouns with their modifiers/complements), VPVerb phrase, PP-Prepositional phrase, AP-Adjectival phrase,
AdvP-Adverbial phrase, FP-Functional phrase (used for secondary predicates).
6
Unless the node it marked [+number], as we will see
shortly.
7
See Fontenelle (1999) for a detailed account of transparent nouns.
3.2
Complex Collocations
As stated, for instance, by (Heid, 1994), collocations can involve more than two (main) terms and
it is possible to adopt a recursive definition of collocations, i.e., complex collocations can be viewed as collocations of collocations. The collocation identification procedure has been extended to
handle such cases. Consider examples (6) below.
(6)a. La voiture tombera probablement en panne
d’essence.
“the car will probably run out of gas”
31
b. natural language processing
The lexicon of the parser was kept constant,
which is to say that both versions used the same
lexicon (which contains slightly more than 7500
English collocation entries), only the parsing module handling collocations was different. From
the output of each parser version, we collected
statistics on the number of collocations (present
in the lexicon) that were identified in the test corpus. More precisely, we traversed the output trees
and counted the items that were marked as collocation heads, each time this was the case (note
that an item may participate in several collocations, not only one). Table 1 presents the number of collocations identified, both with respect to
collocation instances and collocation types.
c. He broke a world record.
In the French sentence (6a), panne d’essence
(literally, “breakdown of gas”, “out of gas”) is
a collocation of type Noun+Prep+Noun, which
combines with the verb tomber (literally, “to
fall”) to form a larger collocation of type
Verb+PrepObject tomber en panne d’essence (“to
run out of gas”). Given the strict left to right
processing order assumed by the parser, it will
first identify the collocation tomber en panne (“to
break down”) when attaching the word panne.
Then, reading the last word, essence (“gas”), the
parser will first identify the collocation panne
d’essence. Since that collocation bears the lexical feature [+partOfCollocation], the identification procedure goes on, through the governors
of that item. The search succeeds with the verb
tomber, and the collocation tomber en panne
d’essence (“run out of gas”) is identified.
4
Tokens
Types
V2
5412
1301
common
4347
1182
V1 only
399
143
V2 only
1003
368
Table 1. Collocation identification results.
As the results show, the new method (column
V2) is more efficient in retrieving collocation instances. It detects 696 more instances, which correspond to an increase of 14.8% relative to the
previous method (column V1). As we lack the
means to compare on a large scale the corresponding syntactic trees, we can only speculate that the
increase is mainly due to the fact that more appropriate analyses are produced by the new method.
A large number of instances are found by both
versions of the parser. The difference between
the two methods is more visible for some syntactic types than for others. Table 2 details the
number of instances of each syntactic type which
are retrieved exclusively by one method or by the
other.
To measure the precision of the two methods,
we randomly selected 20 collocation instances
among those identified by each version of the parser, V1 and V2, and manually checked whether
these instances are correct. Correctness means
that in the given context (i.e., the sentence in
which they were identified), the word combination marked as instance of a lexicalized collocation is indeed an instance of that collocation. A
counterexample would be, for instance, to mark
the pair decision - make in the sentence in (7) as
Evaluation Experiments
In this section, we describe the experiments we
performed in order to evaluate the precision and
recall of the method introduced in section 3, and
to compare it against the previous method (fully
described in Wehrli et al. (2009b)). We extend
this comparison by performing a task-based evaluation, which investigates the impact that the new
method has on the quality of translations produced by a machine translation system relying on
our parser (Wehrli et al., 2009a).
4.1
V1
4716
1218
Precision Evaluation
The data considered in this experiment consist of
a subpart of a corpus of newspaper articles collected from the on-line version of The Economist8 ,
containing slightly more that 0.5 million words.
On these data, we run two versions of our parser:
• V1: a version implementing the previous method of collocation identification,
• V2: a version implementing the new method
described in section 3.
8
URL:http://www.economist.com/
(accessed June, 2010).
32
Syntactic type
A-N
N-N
V-O
V-P-N
N-P-N
V-A
P-N
N&N
Adv-Adv
V1
72
63
22
6
1
25
200
6
4
V2
152
270
190
10
62
166
142
2
9
Difference V2-V1
80
207
168
4
61
141
-58
-4
5
b. Industrial labour costs in western Germany
are higher than in any other country.
To better pinpoint the difference between V1
and V2, we performed a similar evaluation on an
additional set of 20 instances, randomly selected
among the collocations identified exclusively by
each method. Thus, the precision of V1, when
measured on the tokens in ”V1 only”, was 65%.
The precision of V2 on ”V2 only” was 90%. The
2 errors of V2 concern the pair in country, found
in contexts similar to the one shown in example
(8b). The errors of V1 also concerned the same
pair, with one exception – the identification of the
collocation world trade from the context the destruction of the World Trade Centre. Since World
Trade Centre is not in the parser lexicon, V1 analysed it and assigned the first two words to the entry world trade. World was wrongly attached to
Trade, rather than to Centre.
When reported on the totality of the instances
tested, the precision of V1 is 77.5% and that of
V2 is 95%. Besides the increase in the precision
of identified collocations, the new method also
contributes to an increase in the parser coverage10 ,
from 81.7% to 83.3%. The V1 parser version succeeds in building a complete parse tree for 23187
of the total 28375 sentences in the corpus, while
V2 does so for 23629 sentences.
Table 2. Differences between the two methods:
number of tokens retrieved exclusively by each
method.
an instance of the verb-object collocation to make
a decision, which is an entry in our lexicon.
(7)a. The decision to make an offer to buy or sell
property at price is a management decision
that cannot be delegated to staff.
Since judging the correctness of a collocation instance in context is a rather straightforward task,
we do not require multiple judges for this evaluation. The precision obtained is 90% for V1, and
100% for V2.
The small size of test set is motivated by the
fact that the precision is expected to be very high,
since the presence of both collocation components
in a sentence in the relevant syntactic relation almost certainly means that the recognition of the
corresponding collocation is justified. Exceptions
would correspond to a minority of cases in which
the parser either wrongly establishes a relation
between two items which happen to belong to an
entry in the lexicon, or the two items are related
but the combination corresponds to a literal usage
(examples are provided later in this section).
The errors of V1 correspond, in fact, to cases in
which a combination of words used literally was
wrongly attributed to a collocation: in example
(8a), V1 assigned the words on and business to
the lexical entry on business, and in example (8b),
it assigned in and country to the entry in the country9 .
(8)a. It is not, by any means, specific to the
countryside, but it falls especially heavily on
small businesses.
4.2
Recall Evaluation
To compare the recall of two methods we performed a similar experiment, in which we run the two
versions of the parser, V1 and V2, on a small collection of sentences containing annotated collocation instances. These sentences were randomly
selected from the Europarl corpus (Koehn, 2005).
The collocations they contain are all verb-object
collocations. We limit our present investigation
to this syntactic type for two reasons: a) annotating a corpus with all instances of collocation
entries in the lexicon would be a time-consuming
task; and b) verb-object collocations are among
the most syntactically flexible and therefore difficult to detect in real texts. Thus, this test set provides realistic information on recall.
9
V1 makes the same error on (8a), but does better on (8b).
These expressions are frozen and should not be treated as
standard collocations.
10
Coverage refers more precisely to the ratio of sentences
for which a complete parse tree could be built.
33
The test set is divided in two parts: 100 sentences are in English, and 100 other in Italian,
which allows for a cross-linguistic evaluation of
the two methods. Each sentence contains one annotated collocation instance, and there are 10 instances for a collocation type. Table 3 lists the collocation types in the test set (the even rows in column 2 display the glosses for the words in the
Italian collocations).
English
bridge gap
draw distinction
foot bill
give support
hold presidency
meet condition
pose threat
reach compromise
shoulder responsibility
strike balance
English
Italian
V1
63
44
V2
76
66
Common
61
42
V1 only
2
2
V2 only
15
24
Table 4. Recall evaluation results: number of correct collocation instances identified.
4.3
Task-based Evaluation
In addition to reporting the performance results by
using the standard measures of precision and recall, we performed a task-based performance evaluation, in which we quantified the impact that the
newly-proposed method has on the quality of the
output of a machine translation system. As the
examples in table 3 suggest, a literal translation of
collocations is rarely the most appropriate. In fact,
as stated by Orliac and Dillinger (2003), knowledge of collocations is crucial for machine translation systems. An important purpose in identifying collocations with our parser is to enable
their proper treatment in our translation system, a
rule-based system that performs syntactic transfer
by relying on the structures produced by the parser.
In this system, the translation of a collocation
takes place as follows. When the parser identifies a collocation in the source sentence, its component words are marked as collocation members, in order to prevent their literal translation.
When the transfer module processes the collocation head, the system checks in the bilingual
lexicon whether an entry exists for that collocation. If not, the literal translation will apply;
otherwise, the transfer module projects a targetlanguage structure as specified in the corresponding target lexical entry. More precisely, the transfer yields a target language abstract representation, to which grammatical transformations and
morphological generation will apply to create the
target sentence. The identification of collocations
in the source text is a necessary, yet not a sufficient
condition for their successful translation.
In this experiment, we considered the test set
described in section 4.2 and we manually evaluated the translation obtained for each collocation instance. Both subsets (100 English sentences and 100 Italian sentences) were translated
into French. We compared the translations obtai-
Italian
assumere atteggiamento
‘assume’ ‘attitude’
attuare politica
‘carry out’ ‘policy’
avanzare proposta
‘advance’ ‘proposal’
avviare dialogo
‘start’ ‘dialogue’
compiere sforzo
‘commit’ ‘effort’
dare contributo
‘give’ ‘contribution’
dedicare attenzione
‘dedicate’ ‘attention’
operare scelta
‘operate’ ‘choice’
porgere benvenuto
‘give’ ‘welcome’
raggiungere intesa
‘reach’ ‘understanding’
Table 3. Collocation types in the test set.
The evaluation results are presented in table 4.
V1 achieves 63% recall performance on the English data, and 44% on the Italian data. V2 shows
considerably better results: 76% on English and
66% on Italian data. The poorer performance
of both methods on Italian data is explained by
the difference in performance between the English
and Italian parsers, and more precisely, by the difference in their grammatical coverage. The English parser succeeds in building a complete parse
tree for more than 70% of the sentences in the test
set, while the Italian parser only for about 60%.
As found in the previous experiment (presented in section 4.1), for both languages considered
in this experiment, the new method of processing
collocations contributes to improving the parsing
coverage. The coverage of the English parser increases from 71% to 76%, and that of the Italian
parser from 57% to 61%.
34
Task
Collocation identification
Measure
precision
recall
Collocation translation
precision
Parsing
coverage
Test set
40 instances
200 instances
100 instances
100 instances
200 instances
100 instances
100 instances
28375 sentences
200 sentences
200 sentences
Language
English
English, Italian
English
Italian
{English, Italian}-French
English-French
Italian-French
English
English
Italian
Increase
17.5%
17.5%
13%
22%
13%
10%
16%
1.6%
5%
4%
Table 5. Summary of evaluation results.
The experimental results performed showed
that the proposed method, which couples parsing
and collocation identification, leads to substantial improvements in terms of precision and recall over the standard identification method, while
contributing to augment the coverage of the parser. In addition, it was shown that it has a positive impact on the results of a subsequent application, namely, machine translation. Future work
will concentrate on improving our method so that
it accounts for all the possible syntactic configurations of collocational attachments, and on extending its recall evaluation to other syntactic types.
ned by relying on the versions V1 and V2 of our
parser (recall that V2 corresponds to the newlyproposed method and V1 to the previous method).
The use of automatic metrics for evaluating the
translation output was not considered appropriate
in this context, since such n-gram based metrics
underestimate the effect that the substitution of a
single word (like in our case, the verb in a verbobject collocation) has on the fluency, adequacy,
and even on the interpretability of the output sentence.
The comparison showed that, for both language
pairs considered (English-French and ItalianFrench), the version of parser which integrates the
new method is indeed more useful for the machine translation system than the previous version.
When V2 was used, 10 more collocation instances
were correctly translated from English to French
than when using V1. For the Italian-French pair,
V2 helped correctly translating 16 more collocation instances in comparison with V1. This corresponds to an increase in precision of 13% on the
whole test set of 200 sentences. The increase in
performance obtained in all the experiments described in this section is summarized in table 5.
5
Acknowledgements
Thanks to Lorenza Russo and Paola Merlo for a
thorough reading and comments. Part of the research described in this paper has been supported
by a grant from the Swiss National Science Foundation, grant no 100015-117944.
References
Alegria, Iñaki, Olatz Ansa, Xabier Artola, Nerea
Ezeiza, Koldo Gojenola, and Ruben Urizar. 2004.
Representation and treatment of multiword expressions in basque. In Second ACL Workshop on Multiword Expressions: Integrating Processing, pages
48–55, Barcelona, Spain.
Conclusion
In this paper, we addressed the issue of the interconnection between collocation identification and
syntactic parsing, and we proposed an original solution for identifying collocations in a sentence as
soon as possible during the analysis (rather than at
the end of the parsing process). The major advantage of this approach is that collocational information may be used to guide the parser through the
maze of alternatives.
Alshawi, Hiyan and David Carter. 1994. Training
and scaling preference functions for disambiguation. Computational Linguistics, 20(4):635–648.
Benson, Morton, Evelyn Benson, and Robert Ilson.
1986. The BBI Dictionary of English Word Combinations. John Benjamins, Amsterdam/Philadelphia.
Berthouzoz, Cathy and Paola Merlo. 1997. Statistical ambiguity resolution for principle-based parsing. In Nicolov, Nicolas and Ruslan Mitkov, edi-
35
Ratnaparkhi, Adwait. 1998. Statistical models for unsupervised prepositional phrase attachment. In Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics,
pages 1079–1085, Montreal, Quebec, Canada.
tors, Recent Advances in Natural Language Processing: Selected Papers from RANLP’97, Current
Issues in Linguistic Theory, pages 179–186. John
Benjamins, Amsterdam/Philadelphia.
Brun, Caroline. 1998. Terminology finite-state preprocessing for computational LFG. In Proceedings
of the 36th Annual Meeting of the Association for
Computational Linguistics and 17th International
Conference on Computational Linguistics, pages
196–200, Morristown, NJ, USA.
Sag, Ivan A., Timothy Baldwin, Francis Bond, Ann
Copestake, and Dan Flickinger. 2002. Multiword
expressions: A pain in the neck for NLP. In Proceedings of the Third International Conference on
Intelligent Text Processing and Computational Linguistics (CICLING 2002), pages 1–15, Mexico City.
Firth, John R. 1957. Papers in Linguistics 1934-1951.
Oxford Univ. Press, Oxford.
Sinclair, John. 1991. Corpus, Concordance, Collocation. Oxford University Press, Oxford.
Fontenelle, Thierry. 1999. Semantic resources for
word sense disambiguation: a sine qua non? Linguistica e Filologia, (9):25–43. Dipartimento di
Linguistica e Letterature Comparate, Università degli Studi di Bergamo.
Villavicencio, Aline, Valia Kordoni, Yi Zhang, Marco
Idiart, and Carlos Ramisch. 2007. Validation and
evaluation of automatically acquired multiword expressions for grammar engineering. In Proceedings
of the 2007 Joint Conference on Empirical Methods
in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL),
pages 1034–1043, Prague, Czech Republic, June.
Goldman, Jean-Philippe, Luka Nerima, and Eric
Wehrli. 2001. Collocation extraction using a syntactic parser. In Proceedings of the ACL Workshop on Collocation: Computational Extraction,
Analysis and Exploitation, pages 61–66, Toulouse,
France.
Volk, Martin. 2002. Combining unsupervised and
supervised methods for PP attachment disambiguation. In Proceedings of the 19th International Conference on Computational Linguistics (COLING’02), pages 25–32, Taipei, Taiwan.
Grégoire, Nicole, Stefan Evert, and Brigitte Krenn,
editors. 2008. Proceedings of the LREC Workshop
Towards a Shared Task for Multiword Expressions
(MWE 2008). European Language Resources Association (ELRA), Marrakech, Morocco.
Wehrli, Eric and Luka Nerima. 2009. L’analyseur
syntaxique Fips. In Proceedings of the IWPT 2009
ATALA Workshop: What French parsing systems?,
Paris, France.
Heid, Ulrich. 1994. On ways words work together
– research topics in lexical combinatorics. In Proceedings of the 6th Euralex International Congress
on Lexicography (EURALEX ’94), pages 226–257,
Amsterdam, The Netherlands.
Wehrli, Eric, Luka Nerima, and Yves Scherrer. 2009a.
Deep linguistic multilingual translation and bilingual dictionaries. In Proceedings of the Fourth
Workshop on Statistical Machine Translation, pages
90–94, Athens, Greece. Association for Computational Linguistics.
Hindle, Donald and Mats Rooth. 1993. Structural ambiguity and lexical relations. Computational Linguistics, 19(1):103–120.
Wehrli, Eric, Violeta Seretan, Luka Nerima, and Lorenza Russo. 2009b. Collocations in a rule-based
MT system: A case study evaluation of their translation adequacy. In Proceedings of the 13th Annual
Meeting of the European Association for Machine
Translation, pages 128–135, Barcelona, Spain.
Koehn, Philipp. 2005. Europarl: A parallel corpus for
statistical machine translation. In Proceedings of
The Tenth Machine Translation Summit (MT Summit X), pages 79–86, Phuket, Thailand, September.
Wehrli, Eric. 2000. Parsing and collocations. In
Christodoulakis, D., editor, Natural Language Processing, pages 272–282. Springer Verlag.
Orliac, Brigitte and Mike Dillinger. 2003. Collocation
extraction for machine translation. In Proceedings
of Machine Translation Summit IX, pages 292–298,
New Orleans, Lousiana, USA.
Wehrli, Eric. 2007. Fips, a “deep” linguistic multilingual parser. In ACL 2007 Workshop on Deep Linguistic Processing, pages 120–127, Prague, Czech
Republic.
Pantel, Patrick and Dekang Lin. 2000. An unsupervised approach to prepositional phrase attachment
using contextually similar words. In Proceedings
of the 38th Annual Meeting of the Association for
Computational Linguistics, pages 101–108, Hong
Kong, China.
Zhang, Yi and Valia Kordoni. 2006. Automated deep
lexical acquisition for robust open texts processing.
In Proceedings of LREC-2006, pages 275–280, Genoa, Italy.
36